Shared Diffusion Model Loader (Inspire)

This is the Shared Checkpoint Loader's sibling for the modern loading pattern - the one where the diffusion model, text encoders, and VAE come as separate files instead of one fused checkpoint. That's how Flux, SD3, and most of the newer bases ship. So where the vanilla node is "Load Diffusion Model" (the UNET loader), this is that node with Inspire's backend cache bolted on, so the model doesn't get re-read from disk every time you switch workflows.

If you run Flux or another split-file base and hop between graphs, you already know the drag: a 12GB diffusion model reloading on every switch is a real chunk of your iteration time. Cache it once, and subsequent loads are instant. Simple idea, meaningful payoff when you're moving fast.

How it works

Inspire keeps a backend cache - a key-value store living in the ComfyUI process between runs. Load a diffusion model through this node and it stashes the loaded weights under a cache key. Ask for the same model again and it returns the cached copy instead of loading from disk. It stays resident as long as the server is up.

The inputs and outputs that matter

• model_name - the diffusion model file to load, same list as the standard UNET loader.
• weight_dtype (default / fp8_e4m3fn / fp8_e4m3fn_fast / fp8_e5m2) - how to load the weights. default keeps the model's native precision; the fp8 options cast to 8-bit to save VRAM, which is how a lot of people fit Flux on consumer cards. fp8_e4m3fn is the common choice; fp8_e4m3fn_fast trades a hair of quality for speed on supported GPUs. Use default if you have the VRAM and want maximum fidelity.
• key_opt - optional custom cache key. Blank means it keys on model_name, which is fine almost always.
• mode (Auto / Override Cache / Read Only) - Auto for normal caching, Override Cache to force a fresh load and replace the cached copy, Read Only to reuse without creating a new entry.

Outputs: model (wire it to your sampler or model-patch chain like any diffusion model) and a cache key string you can hand to Remove Backend Data (Inspire) to evict it later.

Installing it

Part of the Inspire Pack by Dr.Lt.Data - the ComfyUI Manager and Impact Pack author, so it's first-class tooling. Via ComfyUI Manager: search ComfyUI Inspire Pack, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

then restart. No extra dependencies for this node - it's backend plumbing, not a preprocessor.

Common issues

VRAM stays pinned high. Expected. A cached model is resident until evicted - that's the trade you signed up for. If you cache several big diffusion models in a session, they all stick around. Use the cache key output with Remove Backend Data (Inspire) to clear ones you're done with, or restart to flush.

Changed the file, still getting the old weights. The cache is serving the stale copy. Run once with mode set to Override Cache, then flip back to Auto.

Out of memory even at fp8. Caching doesn't reduce a model's footprint, it keeps it loaded. If you're tight on VRAM, either don't cache this one (use the plain loader) or evict other cached items first. fp8 helps the base cost, but two cached fp8 Flux models still add up.

Nodes won't load after install. Check the terminal log - the usual Inspire Pack cause is a version mismatch. Update Inspire Pack and Impact Pack together and restart.